1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
16 #include "ARMAddressingModes.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMISelLowering.h"
19 #include "ARMMachineFunctionInfo.h"
20 #include "ARMPerfectShuffle.h"
21 #include "ARMRegisterInfo.h"
22 #include "ARMSubtarget.h"
23 #include "ARMTargetMachine.h"
24 #include "ARMTargetObjectFile.h"
25 #include "llvm/CallingConv.h"
26 #include "llvm/Constants.h"
27 #include "llvm/Function.h"
28 #include "llvm/Instruction.h"
29 #include "llvm/Intrinsics.h"
30 #include "llvm/GlobalValue.h"
31 #include "llvm/CodeGen/CallingConvLower.h"
32 #include "llvm/CodeGen/MachineBasicBlock.h"
33 #include "llvm/CodeGen/MachineFrameInfo.h"
34 #include "llvm/CodeGen/MachineFunction.h"
35 #include "llvm/CodeGen/MachineInstrBuilder.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/PseudoSourceValue.h"
38 #include "llvm/CodeGen/SelectionDAG.h"
39 #include "llvm/Target/TargetOptions.h"
40 #include "llvm/ADT/VectorExtras.h"
41 #include "llvm/Support/ErrorHandling.h"
42 #include "llvm/Support/MathExtras.h"
45 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
46 CCValAssign::LocInfo &LocInfo,
47 ISD::ArgFlagsTy &ArgFlags,
49 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
50 CCValAssign::LocInfo &LocInfo,
51 ISD::ArgFlagsTy &ArgFlags,
53 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
54 CCValAssign::LocInfo &LocInfo,
55 ISD::ArgFlagsTy &ArgFlags,
57 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
58 CCValAssign::LocInfo &LocInfo,
59 ISD::ArgFlagsTy &ArgFlags,
62 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
63 EVT PromotedBitwiseVT) {
64 if (VT != PromotedLdStVT) {
65 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
66 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
67 PromotedLdStVT.getSimpleVT());
69 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
70 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
71 PromotedLdStVT.getSimpleVT());
74 EVT ElemTy = VT.getVectorElementType();
75 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
76 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
77 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
78 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
79 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
80 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
81 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Custom);
82 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
84 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
85 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
86 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
89 // Promote all bit-wise operations.
90 if (VT.isInteger() && VT != PromotedBitwiseVT) {
91 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
92 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
93 PromotedBitwiseVT.getSimpleVT());
94 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
95 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
96 PromotedBitwiseVT.getSimpleVT());
97 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
98 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
99 PromotedBitwiseVT.getSimpleVT());
103 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
104 addRegisterClass(VT, ARM::DPRRegisterClass);
105 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
108 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
109 addRegisterClass(VT, ARM::QPRRegisterClass);
110 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
113 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
114 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
115 return new TargetLoweringObjectFileMachO();
116 return new ARMElfTargetObjectFile();
119 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
120 : TargetLowering(TM, createTLOF(TM)), ARMPCLabelIndex(0) {
121 Subtarget = &TM.getSubtarget<ARMSubtarget>();
123 if (Subtarget->isTargetDarwin()) {
124 // Uses VFP for Thumb libfuncs if available.
125 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
126 // Single-precision floating-point arithmetic.
127 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
128 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
129 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
130 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
132 // Double-precision floating-point arithmetic.
133 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
134 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
135 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
136 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
138 // Single-precision comparisons.
139 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
140 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
141 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
142 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
143 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
144 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
145 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
146 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
148 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
149 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
150 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
151 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
152 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
153 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
154 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
155 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
157 // Double-precision comparisons.
158 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
159 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
160 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
161 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
162 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
163 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
164 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
165 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
167 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
168 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
169 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
170 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
171 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
172 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
173 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
174 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
176 // Floating-point to integer conversions.
177 // i64 conversions are done via library routines even when generating VFP
178 // instructions, so use the same ones.
179 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
180 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
181 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
182 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
184 // Conversions between floating types.
185 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
186 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
188 // Integer to floating-point conversions.
189 // i64 conversions are done via library routines even when generating VFP
190 // instructions, so use the same ones.
191 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
192 // e.g., __floatunsidf vs. __floatunssidfvfp.
193 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
194 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
195 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
196 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
200 // These libcalls are not available in 32-bit.
201 setLibcallName(RTLIB::SHL_I128, 0);
202 setLibcallName(RTLIB::SRL_I128, 0);
203 setLibcallName(RTLIB::SRA_I128, 0);
205 // Libcalls should use the AAPCS base standard ABI, even if hard float
206 // is in effect, as per the ARM RTABI specification, section 4.1.2.
207 if (Subtarget->isAAPCS_ABI()) {
208 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
209 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
210 CallingConv::ARM_AAPCS);
214 if (Subtarget->isThumb1Only())
215 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
217 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
218 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
219 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
220 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
222 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
225 if (Subtarget->hasNEON()) {
226 addDRTypeForNEON(MVT::v2f32);
227 addDRTypeForNEON(MVT::v8i8);
228 addDRTypeForNEON(MVT::v4i16);
229 addDRTypeForNEON(MVT::v2i32);
230 addDRTypeForNEON(MVT::v1i64);
232 addQRTypeForNEON(MVT::v4f32);
233 addQRTypeForNEON(MVT::v2f64);
234 addQRTypeForNEON(MVT::v16i8);
235 addQRTypeForNEON(MVT::v8i16);
236 addQRTypeForNEON(MVT::v4i32);
237 addQRTypeForNEON(MVT::v2i64);
239 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
240 setTargetDAGCombine(ISD::SHL);
241 setTargetDAGCombine(ISD::SRL);
242 setTargetDAGCombine(ISD::SRA);
243 setTargetDAGCombine(ISD::SIGN_EXTEND);
244 setTargetDAGCombine(ISD::ZERO_EXTEND);
245 setTargetDAGCombine(ISD::ANY_EXTEND);
248 computeRegisterProperties();
250 // ARM does not have f32 extending load.
251 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
253 // ARM does not have i1 sign extending load.
254 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
256 // ARM supports all 4 flavors of integer indexed load / store.
257 if (!Subtarget->isThumb1Only()) {
258 for (unsigned im = (unsigned)ISD::PRE_INC;
259 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
260 setIndexedLoadAction(im, MVT::i1, Legal);
261 setIndexedLoadAction(im, MVT::i8, Legal);
262 setIndexedLoadAction(im, MVT::i16, Legal);
263 setIndexedLoadAction(im, MVT::i32, Legal);
264 setIndexedStoreAction(im, MVT::i1, Legal);
265 setIndexedStoreAction(im, MVT::i8, Legal);
266 setIndexedStoreAction(im, MVT::i16, Legal);
267 setIndexedStoreAction(im, MVT::i32, Legal);
271 // i64 operation support.
272 if (Subtarget->isThumb1Only()) {
273 setOperationAction(ISD::MUL, MVT::i64, Expand);
274 setOperationAction(ISD::MULHU, MVT::i32, Expand);
275 setOperationAction(ISD::MULHS, MVT::i32, Expand);
276 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
277 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
279 setOperationAction(ISD::MUL, MVT::i64, Expand);
280 setOperationAction(ISD::MULHU, MVT::i32, Expand);
281 if (!Subtarget->hasV6Ops())
282 setOperationAction(ISD::MULHS, MVT::i32, Expand);
284 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
285 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
286 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
287 setOperationAction(ISD::SRL, MVT::i64, Custom);
288 setOperationAction(ISD::SRA, MVT::i64, Custom);
290 // ARM does not have ROTL.
291 setOperationAction(ISD::ROTL, MVT::i32, Expand);
292 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
293 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
294 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
295 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
297 // Only ARMv6 has BSWAP.
298 if (!Subtarget->hasV6Ops())
299 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
301 // These are expanded into libcalls.
302 setOperationAction(ISD::SDIV, MVT::i32, Expand);
303 setOperationAction(ISD::UDIV, MVT::i32, Expand);
304 setOperationAction(ISD::SREM, MVT::i32, Expand);
305 setOperationAction(ISD::UREM, MVT::i32, Expand);
306 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
307 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
309 // Support label based line numbers.
310 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
311 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
313 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
314 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
315 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
316 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
318 // Use the default implementation.
319 setOperationAction(ISD::VASTART, MVT::Other, Custom);
320 setOperationAction(ISD::VAARG, MVT::Other, Expand);
321 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
322 setOperationAction(ISD::VAEND, MVT::Other, Expand);
323 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
324 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
325 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
326 // FIXME: Shouldn't need this, since no register is used, but the legalizer
327 // doesn't yet know how to not do that for SjLj.
328 setExceptionSelectorRegister(ARM::R0);
329 if (Subtarget->isThumb())
330 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
332 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
333 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
335 if (!Subtarget->hasV6Ops() && !Subtarget->isThumb2()) {
336 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
337 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
339 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
341 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
342 // Turn f64->i64 into FMRRD, i64 -> f64 to FMDRR iff target supports vfp2.
343 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
345 // We want to custom lower some of our intrinsics.
346 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
347 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
348 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
350 setOperationAction(ISD::SETCC, MVT::i32, Expand);
351 setOperationAction(ISD::SETCC, MVT::f32, Expand);
352 setOperationAction(ISD::SETCC, MVT::f64, Expand);
353 setOperationAction(ISD::SELECT, MVT::i32, Expand);
354 setOperationAction(ISD::SELECT, MVT::f32, Expand);
355 setOperationAction(ISD::SELECT, MVT::f64, Expand);
356 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
357 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
358 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
360 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
361 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
362 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
363 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
364 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
366 // We don't support sin/cos/fmod/copysign/pow
367 setOperationAction(ISD::FSIN, MVT::f64, Expand);
368 setOperationAction(ISD::FSIN, MVT::f32, Expand);
369 setOperationAction(ISD::FCOS, MVT::f32, Expand);
370 setOperationAction(ISD::FCOS, MVT::f64, Expand);
371 setOperationAction(ISD::FREM, MVT::f64, Expand);
372 setOperationAction(ISD::FREM, MVT::f32, Expand);
373 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
374 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
375 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
377 setOperationAction(ISD::FPOW, MVT::f64, Expand);
378 setOperationAction(ISD::FPOW, MVT::f32, Expand);
380 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
381 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
382 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
383 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
384 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
385 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
388 // We have target-specific dag combine patterns for the following nodes:
389 // ARMISD::FMRRD - No need to call setTargetDAGCombine
390 setTargetDAGCombine(ISD::ADD);
391 setTargetDAGCombine(ISD::SUB);
393 setStackPointerRegisterToSaveRestore(ARM::SP);
394 setSchedulingPreference(SchedulingForRegPressure);
396 // FIXME: If-converter should use instruction latency to determine
397 // profitability rather than relying on fixed limits.
398 if (Subtarget->getCPUString() == "generic") {
399 // Generic (and overly aggressive) if-conversion limits.
400 setIfCvtBlockSizeLimit(10);
401 setIfCvtDupBlockSizeLimit(2);
402 } else if (Subtarget->hasV6Ops()) {
403 setIfCvtBlockSizeLimit(2);
404 setIfCvtDupBlockSizeLimit(1);
406 setIfCvtBlockSizeLimit(3);
407 setIfCvtDupBlockSizeLimit(2);
410 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
411 // Do not enable CodePlacementOpt for now: it currently runs after the
412 // ARMConstantIslandPass and messes up branch relaxation and placement
413 // of constant islands.
414 // benefitFromCodePlacementOpt = true;
417 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
420 case ARMISD::Wrapper: return "ARMISD::Wrapper";
421 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
422 case ARMISD::CALL: return "ARMISD::CALL";
423 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
424 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
425 case ARMISD::tCALL: return "ARMISD::tCALL";
426 case ARMISD::BRCOND: return "ARMISD::BRCOND";
427 case ARMISD::BR_JT: return "ARMISD::BR_JT";
428 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
429 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
430 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
431 case ARMISD::CMP: return "ARMISD::CMP";
432 case ARMISD::CMPZ: return "ARMISD::CMPZ";
433 case ARMISD::CMPFP: return "ARMISD::CMPFP";
434 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
435 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
436 case ARMISD::CMOV: return "ARMISD::CMOV";
437 case ARMISD::CNEG: return "ARMISD::CNEG";
439 case ARMISD::FTOSI: return "ARMISD::FTOSI";
440 case ARMISD::FTOUI: return "ARMISD::FTOUI";
441 case ARMISD::SITOF: return "ARMISD::SITOF";
442 case ARMISD::UITOF: return "ARMISD::UITOF";
444 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
445 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
446 case ARMISD::RRX: return "ARMISD::RRX";
448 case ARMISD::FMRRD: return "ARMISD::FMRRD";
449 case ARMISD::FMDRR: return "ARMISD::FMDRR";
451 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
453 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
455 case ARMISD::VCEQ: return "ARMISD::VCEQ";
456 case ARMISD::VCGE: return "ARMISD::VCGE";
457 case ARMISD::VCGEU: return "ARMISD::VCGEU";
458 case ARMISD::VCGT: return "ARMISD::VCGT";
459 case ARMISD::VCGTU: return "ARMISD::VCGTU";
460 case ARMISD::VTST: return "ARMISD::VTST";
462 case ARMISD::VSHL: return "ARMISD::VSHL";
463 case ARMISD::VSHRs: return "ARMISD::VSHRs";
464 case ARMISD::VSHRu: return "ARMISD::VSHRu";
465 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
466 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
467 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
468 case ARMISD::VSHRN: return "ARMISD::VSHRN";
469 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
470 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
471 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
472 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
473 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
474 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
475 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
476 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
477 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
478 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
479 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
480 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
481 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
482 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
483 case ARMISD::VDUP: return "ARMISD::VDUP";
484 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
485 case ARMISD::VEXT: return "ARMISD::VEXT";
486 case ARMISD::VREV64: return "ARMISD::VREV64";
487 case ARMISD::VREV32: return "ARMISD::VREV32";
488 case ARMISD::VREV16: return "ARMISD::VREV16";
489 case ARMISD::VZIP: return "ARMISD::VZIP";
490 case ARMISD::VUZP: return "ARMISD::VUZP";
491 case ARMISD::VTRN: return "ARMISD::VTRN";
495 /// getFunctionAlignment - Return the Log2 alignment of this function.
496 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
497 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
500 //===----------------------------------------------------------------------===//
502 //===----------------------------------------------------------------------===//
504 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
505 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
507 default: llvm_unreachable("Unknown condition code!");
508 case ISD::SETNE: return ARMCC::NE;
509 case ISD::SETEQ: return ARMCC::EQ;
510 case ISD::SETGT: return ARMCC::GT;
511 case ISD::SETGE: return ARMCC::GE;
512 case ISD::SETLT: return ARMCC::LT;
513 case ISD::SETLE: return ARMCC::LE;
514 case ISD::SETUGT: return ARMCC::HI;
515 case ISD::SETUGE: return ARMCC::HS;
516 case ISD::SETULT: return ARMCC::LO;
517 case ISD::SETULE: return ARMCC::LS;
521 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC. It
522 /// returns true if the operands should be inverted to form the proper
524 static bool FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
525 ARMCC::CondCodes &CondCode2) {
527 CondCode2 = ARMCC::AL;
529 default: llvm_unreachable("Unknown FP condition!");
531 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
533 case ISD::SETOGT: CondCode = ARMCC::GT; break;
535 case ISD::SETOGE: CondCode = ARMCC::GE; break;
536 case ISD::SETOLT: CondCode = ARMCC::MI; break;
537 case ISD::SETOLE: CondCode = ARMCC::GT; Invert = true; break;
538 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
539 case ISD::SETO: CondCode = ARMCC::VC; break;
540 case ISD::SETUO: CondCode = ARMCC::VS; break;
541 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
542 case ISD::SETUGT: CondCode = ARMCC::HI; break;
543 case ISD::SETUGE: CondCode = ARMCC::PL; break;
545 case ISD::SETULT: CondCode = ARMCC::LT; break;
547 case ISD::SETULE: CondCode = ARMCC::LE; break;
549 case ISD::SETUNE: CondCode = ARMCC::NE; break;
554 //===----------------------------------------------------------------------===//
555 // Calling Convention Implementation
556 //===----------------------------------------------------------------------===//
558 #include "ARMGenCallingConv.inc"
560 // APCS f64 is in register pairs, possibly split to stack
561 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
562 CCValAssign::LocInfo &LocInfo,
563 CCState &State, bool CanFail) {
564 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
566 // Try to get the first register.
567 if (unsigned Reg = State.AllocateReg(RegList, 4))
568 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
570 // For the 2nd half of a v2f64, do not fail.
574 // Put the whole thing on the stack.
575 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
576 State.AllocateStack(8, 4),
581 // Try to get the second register.
582 if (unsigned Reg = State.AllocateReg(RegList, 4))
583 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
585 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
586 State.AllocateStack(4, 4),
591 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
592 CCValAssign::LocInfo &LocInfo,
593 ISD::ArgFlagsTy &ArgFlags,
595 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
597 if (LocVT == MVT::v2f64 &&
598 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
600 return true; // we handled it
603 // AAPCS f64 is in aligned register pairs
604 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
605 CCValAssign::LocInfo &LocInfo,
606 CCState &State, bool CanFail) {
607 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
608 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
610 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
612 // For the 2nd half of a v2f64, do not just fail.
616 // Put the whole thing on the stack.
617 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
618 State.AllocateStack(8, 8),
624 for (i = 0; i < 2; ++i)
625 if (HiRegList[i] == Reg)
628 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
629 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
634 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
635 CCValAssign::LocInfo &LocInfo,
636 ISD::ArgFlagsTy &ArgFlags,
638 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
640 if (LocVT == MVT::v2f64 &&
641 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
643 return true; // we handled it
646 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
647 CCValAssign::LocInfo &LocInfo, CCState &State) {
648 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
649 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
651 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
653 return false; // we didn't handle it
656 for (i = 0; i < 2; ++i)
657 if (HiRegList[i] == Reg)
660 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
661 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
666 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
667 CCValAssign::LocInfo &LocInfo,
668 ISD::ArgFlagsTy &ArgFlags,
670 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
672 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
674 return true; // we handled it
677 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
678 CCValAssign::LocInfo &LocInfo,
679 ISD::ArgFlagsTy &ArgFlags,
681 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
685 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
686 /// given CallingConvention value.
687 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(unsigned CC,
689 bool isVarArg) const {
692 llvm_unreachable("Unsupported calling convention");
694 case CallingConv::Fast:
695 // Use target triple & subtarget features to do actual dispatch.
696 if (Subtarget->isAAPCS_ABI()) {
697 if (Subtarget->hasVFP2() &&
698 FloatABIType == FloatABI::Hard && !isVarArg)
699 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
701 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
703 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
704 case CallingConv::ARM_AAPCS_VFP:
705 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
706 case CallingConv::ARM_AAPCS:
707 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
708 case CallingConv::ARM_APCS:
709 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
713 /// LowerCallResult - Lower the result values of a call into the
714 /// appropriate copies out of appropriate physical registers.
716 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
717 unsigned CallConv, bool isVarArg,
718 const SmallVectorImpl<ISD::InputArg> &Ins,
719 DebugLoc dl, SelectionDAG &DAG,
720 SmallVectorImpl<SDValue> &InVals) {
722 // Assign locations to each value returned by this call.
723 SmallVector<CCValAssign, 16> RVLocs;
724 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
725 RVLocs, *DAG.getContext());
726 CCInfo.AnalyzeCallResult(Ins,
727 CCAssignFnForNode(CallConv, /* Return*/ true,
730 // Copy all of the result registers out of their specified physreg.
731 for (unsigned i = 0; i != RVLocs.size(); ++i) {
732 CCValAssign VA = RVLocs[i];
735 if (VA.needsCustom()) {
736 // Handle f64 or half of a v2f64.
737 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
739 Chain = Lo.getValue(1);
740 InFlag = Lo.getValue(2);
741 VA = RVLocs[++i]; // skip ahead to next loc
742 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
744 Chain = Hi.getValue(1);
745 InFlag = Hi.getValue(2);
746 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
748 if (VA.getLocVT() == MVT::v2f64) {
749 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
750 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
751 DAG.getConstant(0, MVT::i32));
753 VA = RVLocs[++i]; // skip ahead to next loc
754 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
755 Chain = Lo.getValue(1);
756 InFlag = Lo.getValue(2);
757 VA = RVLocs[++i]; // skip ahead to next loc
758 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
759 Chain = Hi.getValue(1);
760 InFlag = Hi.getValue(2);
761 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
762 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
763 DAG.getConstant(1, MVT::i32));
766 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
768 Chain = Val.getValue(1);
769 InFlag = Val.getValue(2);
772 switch (VA.getLocInfo()) {
773 default: llvm_unreachable("Unknown loc info!");
774 case CCValAssign::Full: break;
775 case CCValAssign::BCvt:
776 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
780 InVals.push_back(Val);
786 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
787 /// by "Src" to address "Dst" of size "Size". Alignment information is
788 /// specified by the specific parameter attribute. The copy will be passed as
789 /// a byval function parameter.
790 /// Sometimes what we are copying is the end of a larger object, the part that
791 /// does not fit in registers.
793 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
794 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
796 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
797 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
798 /*AlwaysInline=*/false, NULL, 0, NULL, 0);
801 /// LowerMemOpCallTo - Store the argument to the stack.
803 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
804 SDValue StackPtr, SDValue Arg,
805 DebugLoc dl, SelectionDAG &DAG,
806 const CCValAssign &VA,
807 ISD::ArgFlagsTy Flags) {
808 unsigned LocMemOffset = VA.getLocMemOffset();
809 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
810 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
811 if (Flags.isByVal()) {
812 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
814 return DAG.getStore(Chain, dl, Arg, PtrOff,
815 PseudoSourceValue::getStack(), LocMemOffset);
818 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
819 SDValue Chain, SDValue &Arg,
820 RegsToPassVector &RegsToPass,
821 CCValAssign &VA, CCValAssign &NextVA,
823 SmallVector<SDValue, 8> &MemOpChains,
824 ISD::ArgFlagsTy Flags) {
826 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
827 DAG.getVTList(MVT::i32, MVT::i32), Arg);
828 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
830 if (NextVA.isRegLoc())
831 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
833 assert(NextVA.isMemLoc());
834 if (StackPtr.getNode() == 0)
835 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
837 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
843 /// LowerCall - Lowering a call into a callseq_start <-
844 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
847 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
848 unsigned CallConv, bool isVarArg,
850 const SmallVectorImpl<ISD::OutputArg> &Outs,
851 const SmallVectorImpl<ISD::InputArg> &Ins,
852 DebugLoc dl, SelectionDAG &DAG,
853 SmallVectorImpl<SDValue> &InVals) {
855 // Analyze operands of the call, assigning locations to each operand.
856 SmallVector<CCValAssign, 16> ArgLocs;
857 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
859 CCInfo.AnalyzeCallOperands(Outs,
860 CCAssignFnForNode(CallConv, /* Return*/ false,
863 // Get a count of how many bytes are to be pushed on the stack.
864 unsigned NumBytes = CCInfo.getNextStackOffset();
866 // Adjust the stack pointer for the new arguments...
867 // These operations are automatically eliminated by the prolog/epilog pass
868 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
870 SDValue StackPtr = DAG.getRegister(ARM::SP, MVT::i32);
872 RegsToPassVector RegsToPass;
873 SmallVector<SDValue, 8> MemOpChains;
875 // Walk the register/memloc assignments, inserting copies/loads. In the case
876 // of tail call optimization, arguments are handled later.
877 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
880 CCValAssign &VA = ArgLocs[i];
881 SDValue Arg = Outs[realArgIdx].Val;
882 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
884 // Promote the value if needed.
885 switch (VA.getLocInfo()) {
886 default: llvm_unreachable("Unknown loc info!");
887 case CCValAssign::Full: break;
888 case CCValAssign::SExt:
889 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
891 case CCValAssign::ZExt:
892 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
894 case CCValAssign::AExt:
895 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
897 case CCValAssign::BCvt:
898 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
902 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
903 if (VA.needsCustom()) {
904 if (VA.getLocVT() == MVT::v2f64) {
905 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
906 DAG.getConstant(0, MVT::i32));
907 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
908 DAG.getConstant(1, MVT::i32));
910 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
911 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
913 VA = ArgLocs[++i]; // skip ahead to next loc
915 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
916 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
918 assert(VA.isMemLoc());
919 if (StackPtr.getNode() == 0)
920 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
922 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
923 dl, DAG, VA, Flags));
926 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
927 StackPtr, MemOpChains, Flags);
929 } else if (VA.isRegLoc()) {
930 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
932 assert(VA.isMemLoc());
933 if (StackPtr.getNode() == 0)
934 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
936 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
937 dl, DAG, VA, Flags));
941 if (!MemOpChains.empty())
942 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
943 &MemOpChains[0], MemOpChains.size());
945 // Build a sequence of copy-to-reg nodes chained together with token chain
946 // and flag operands which copy the outgoing args into the appropriate regs.
948 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
949 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
950 RegsToPass[i].second, InFlag);
951 InFlag = Chain.getValue(1);
954 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
955 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
956 // node so that legalize doesn't hack it.
957 bool isDirect = false;
958 bool isARMFunc = false;
959 bool isLocalARMFunc = false;
960 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
961 GlobalValue *GV = G->getGlobal();
963 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
964 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
965 getTargetMachine().getRelocationModel() != Reloc::Static;
966 isARMFunc = !Subtarget->isThumb() || isStub;
967 // ARM call to a local ARM function is predicable.
968 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
969 // tBX takes a register source operand.
970 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
971 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
973 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
974 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
975 Callee = DAG.getLoad(getPointerTy(), dl,
976 DAG.getEntryNode(), CPAddr, NULL, 0);
977 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
978 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
979 getPointerTy(), Callee, PICLabel);
981 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
982 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
984 bool isStub = Subtarget->isTargetDarwin() &&
985 getTargetMachine().getRelocationModel() != Reloc::Static;
986 isARMFunc = !Subtarget->isThumb() || isStub;
987 // tBX takes a register source operand.
988 const char *Sym = S->getSymbol();
989 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
990 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
991 Sym, ARMPCLabelIndex, 4);
992 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
993 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
994 Callee = DAG.getLoad(getPointerTy(), dl,
995 DAG.getEntryNode(), CPAddr, NULL, 0);
996 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
997 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
998 getPointerTy(), Callee, PICLabel);
1000 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1003 // FIXME: handle tail calls differently.
1005 if (Subtarget->isThumb()) {
1006 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1007 CallOpc = ARMISD::CALL_NOLINK;
1009 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1011 CallOpc = (isDirect || Subtarget->hasV5TOps())
1012 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1013 : ARMISD::CALL_NOLINK;
1015 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1016 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1017 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1018 InFlag = Chain.getValue(1);
1021 std::vector<SDValue> Ops;
1022 Ops.push_back(Chain);
1023 Ops.push_back(Callee);
1025 // Add argument registers to the end of the list so that they are known live
1027 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1028 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1029 RegsToPass[i].second.getValueType()));
1031 if (InFlag.getNode())
1032 Ops.push_back(InFlag);
1033 // Returns a chain and a flag for retval copy to use.
1034 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1035 &Ops[0], Ops.size());
1036 InFlag = Chain.getValue(1);
1038 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1039 DAG.getIntPtrConstant(0, true), InFlag);
1041 InFlag = Chain.getValue(1);
1043 // Handle result values, copying them out of physregs into vregs that we
1045 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1050 ARMTargetLowering::LowerReturn(SDValue Chain,
1051 unsigned CallConv, bool isVarArg,
1052 const SmallVectorImpl<ISD::OutputArg> &Outs,
1053 DebugLoc dl, SelectionDAG &DAG) {
1055 // CCValAssign - represent the assignment of the return value to a location.
1056 SmallVector<CCValAssign, 16> RVLocs;
1058 // CCState - Info about the registers and stack slots.
1059 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1062 // Analyze outgoing return values.
1063 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1066 // If this is the first return lowered for this function, add
1067 // the regs to the liveout set for the function.
1068 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1069 for (unsigned i = 0; i != RVLocs.size(); ++i)
1070 if (RVLocs[i].isRegLoc())
1071 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1076 // Copy the result values into the output registers.
1077 for (unsigned i = 0, realRVLocIdx = 0;
1079 ++i, ++realRVLocIdx) {
1080 CCValAssign &VA = RVLocs[i];
1081 assert(VA.isRegLoc() && "Can only return in registers!");
1083 SDValue Arg = Outs[realRVLocIdx].Val;
1085 switch (VA.getLocInfo()) {
1086 default: llvm_unreachable("Unknown loc info!");
1087 case CCValAssign::Full: break;
1088 case CCValAssign::BCvt:
1089 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1093 if (VA.needsCustom()) {
1094 if (VA.getLocVT() == MVT::v2f64) {
1095 // Extract the first half and return it in two registers.
1096 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1097 DAG.getConstant(0, MVT::i32));
1098 SDValue HalfGPRs = DAG.getNode(ARMISD::FMRRD, dl,
1099 DAG.getVTList(MVT::i32, MVT::i32), Half);
1101 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1102 Flag = Chain.getValue(1);
1103 VA = RVLocs[++i]; // skip ahead to next loc
1104 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1105 HalfGPRs.getValue(1), Flag);
1106 Flag = Chain.getValue(1);
1107 VA = RVLocs[++i]; // skip ahead to next loc
1109 // Extract the 2nd half and fall through to handle it as an f64 value.
1110 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1111 DAG.getConstant(1, MVT::i32));
1113 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1115 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
1116 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1117 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1118 Flag = Chain.getValue(1);
1119 VA = RVLocs[++i]; // skip ahead to next loc
1120 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1123 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1125 // Guarantee that all emitted copies are
1126 // stuck together, avoiding something bad.
1127 Flag = Chain.getValue(1);
1132 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1134 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1139 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1140 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1141 // one of the above mentioned nodes. It has to be wrapped because otherwise
1142 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1143 // be used to form addressing mode. These wrapped nodes will be selected
1145 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1146 EVT PtrVT = Op.getValueType();
1147 // FIXME there is no actual debug info here
1148 DebugLoc dl = Op.getDebugLoc();
1149 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1151 if (CP->isMachineConstantPoolEntry())
1152 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1153 CP->getAlignment());
1155 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1156 CP->getAlignment());
1157 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1160 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1162 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1163 SelectionDAG &DAG) {
1164 DebugLoc dl = GA->getDebugLoc();
1165 EVT PtrVT = getPointerTy();
1166 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1167 ARMConstantPoolValue *CPV =
1168 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1169 PCAdj, "tlsgd", true);
1170 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1171 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1172 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument, NULL, 0);
1173 SDValue Chain = Argument.getValue(1);
1175 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1176 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1178 // call __tls_get_addr.
1181 Entry.Node = Argument;
1182 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1183 Args.push_back(Entry);
1184 // FIXME: is there useful debug info available here?
1185 std::pair<SDValue, SDValue> CallResult =
1186 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1187 false, false, false, false,
1188 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1189 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1190 return CallResult.first;
1193 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1194 // "local exec" model.
1196 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1197 SelectionDAG &DAG) {
1198 GlobalValue *GV = GA->getGlobal();
1199 DebugLoc dl = GA->getDebugLoc();
1201 SDValue Chain = DAG.getEntryNode();
1202 EVT PtrVT = getPointerTy();
1203 // Get the Thread Pointer
1204 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1206 if (GV->isDeclaration()) {
1207 // initial exec model
1208 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1209 ARMConstantPoolValue *CPV =
1210 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1211 PCAdj, "gottpoff", true);
1212 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1213 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1214 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1215 Chain = Offset.getValue(1);
1217 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1218 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1220 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1223 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1224 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1225 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1226 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1229 // The address of the thread local variable is the add of the thread
1230 // pointer with the offset of the variable.
1231 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1235 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
1236 // TODO: implement the "local dynamic" model
1237 assert(Subtarget->isTargetELF() &&
1238 "TLS not implemented for non-ELF targets");
1239 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1240 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1241 // otherwise use the "Local Exec" TLS Model
1242 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1243 return LowerToTLSGeneralDynamicModel(GA, DAG);
1245 return LowerToTLSExecModels(GA, DAG);
1248 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1249 SelectionDAG &DAG) {
1250 EVT PtrVT = getPointerTy();
1251 DebugLoc dl = Op.getDebugLoc();
1252 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1253 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1254 if (RelocM == Reloc::PIC_) {
1255 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1256 ARMConstantPoolValue *CPV =
1257 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1258 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1259 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1260 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1262 SDValue Chain = Result.getValue(1);
1263 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1264 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1266 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
1269 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1270 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1271 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1275 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1276 SelectionDAG &DAG) {
1277 EVT PtrVT = getPointerTy();
1278 DebugLoc dl = Op.getDebugLoc();
1279 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1280 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1282 if (RelocM == Reloc::Static)
1283 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1285 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1286 ARMConstantPoolValue *CPV =
1287 new ARMConstantPoolValue(GV, ARMPCLabelIndex, PCAdj);
1288 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1290 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1292 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1293 SDValue Chain = Result.getValue(1);
1295 if (RelocM == Reloc::PIC_) {
1296 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1297 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1300 if (Subtarget->GVIsIndirectSymbol(GV, RelocM == Reloc::Static))
1301 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
1306 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1308 assert(Subtarget->isTargetELF() &&
1309 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1310 EVT PtrVT = getPointerTy();
1311 DebugLoc dl = Op.getDebugLoc();
1312 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1313 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1314 "_GLOBAL_OFFSET_TABLE_",
1315 ARMPCLabelIndex, PCAdj);
1316 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1317 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1318 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1319 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1320 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1323 static SDValue LowerNeonVLDIntrinsic(SDValue Op, SelectionDAG &DAG,
1325 SDNode *Node = Op.getNode();
1326 EVT VT = Node->getValueType(0);
1328 // No expansion needed for 64-bit vectors.
1329 if (VT.is64BitVector())
1332 // FIXME: We need to expand VLD3 and VLD4 of 128-bit vectors into separate
1333 // operations to load the even and odd registers.
1337 static SDValue LowerNeonVSTIntrinsic(SDValue Op, SelectionDAG &DAG,
1339 SDNode *Node = Op.getNode();
1340 EVT VT = Node->getOperand(3).getValueType();
1342 // No expansion needed for 64-bit vectors.
1343 if (VT.is64BitVector())
1346 // FIXME: We need to expand VST3 and VST4 of 128-bit vectors into separate
1347 // operations to store the even and odd registers.
1352 ARMTargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) {
1353 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1355 case Intrinsic::arm_neon_vld3:
1356 return LowerNeonVLDIntrinsic(Op, DAG, 3);
1357 case Intrinsic::arm_neon_vld4:
1358 return LowerNeonVLDIntrinsic(Op, DAG, 4);
1359 case Intrinsic::arm_neon_vst3:
1360 return LowerNeonVSTIntrinsic(Op, DAG, 3);
1361 case Intrinsic::arm_neon_vst4:
1362 return LowerNeonVSTIntrinsic(Op, DAG, 4);
1363 default: return SDValue(); // Don't custom lower most intrinsics.
1368 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
1369 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1370 DebugLoc dl = Op.getDebugLoc();
1372 default: return SDValue(); // Don't custom lower most intrinsics.
1373 case Intrinsic::arm_thread_pointer: {
1374 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1375 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1377 case Intrinsic::eh_sjlj_lsda: {
1378 // blah. horrible, horrible hack with the forced magic name.
1379 // really need to clean this up. It belongs in the target-independent
1380 // layer somehow that doesn't require the coupling with the asm
1382 MachineFunction &MF = DAG.getMachineFunction();
1383 EVT PtrVT = getPointerTy();
1384 DebugLoc dl = Op.getDebugLoc();
1385 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1387 unsigned PCAdj = (RelocM != Reloc::PIC_)
1388 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1389 // Save off the LSDA name for the AsmPrinter to use when it's time
1390 // to emit the table
1391 std::string LSDAName = "L_lsda_";
1392 LSDAName += MF.getFunction()->getName();
1393 ARMConstantPoolValue *CPV =
1394 new ARMConstantPoolValue(*DAG.getContext(), LSDAName.c_str(),
1395 ARMPCLabelIndex, PCAdj);
1396 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1397 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1399 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1400 SDValue Chain = Result.getValue(1);
1402 if (RelocM == Reloc::PIC_) {
1403 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1404 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1408 case Intrinsic::eh_sjlj_setjmp:
1409 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1));
1413 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
1414 unsigned VarArgsFrameIndex) {
1415 // vastart just stores the address of the VarArgsFrameIndex slot into the
1416 // memory location argument.
1417 DebugLoc dl = Op.getDebugLoc();
1418 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1419 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1420 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1421 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
1425 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) {
1426 SDNode *Node = Op.getNode();
1427 DebugLoc dl = Node->getDebugLoc();
1428 EVT VT = Node->getValueType(0);
1429 SDValue Chain = Op.getOperand(0);
1430 SDValue Size = Op.getOperand(1);
1431 SDValue Align = Op.getOperand(2);
1433 // Chain the dynamic stack allocation so that it doesn't modify the stack
1434 // pointer when other instructions are using the stack.
1435 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1437 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1438 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1439 if (AlignVal > StackAlign)
1440 // Do this now since selection pass cannot introduce new target
1441 // independent node.
1442 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1444 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1445 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1446 // do even more horrible hack later.
1447 MachineFunction &MF = DAG.getMachineFunction();
1448 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1449 if (AFI->isThumb1OnlyFunction()) {
1451 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1453 uint32_t Val = C->getZExtValue();
1454 if (Val <= 508 && ((Val & 3) == 0))
1458 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1461 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1462 SDValue Ops1[] = { Chain, Size, Align };
1463 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1464 Chain = Res.getValue(1);
1465 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1466 DAG.getIntPtrConstant(0, true), SDValue());
1467 SDValue Ops2[] = { Res, Chain };
1468 return DAG.getMergeValues(Ops2, 2, dl);
1472 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1473 SDValue &Root, SelectionDAG &DAG,
1475 MachineFunction &MF = DAG.getMachineFunction();
1476 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1478 TargetRegisterClass *RC;
1479 if (AFI->isThumb1OnlyFunction())
1480 RC = ARM::tGPRRegisterClass;
1482 RC = ARM::GPRRegisterClass;
1484 // Transform the arguments stored in physical registers into virtual ones.
1485 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1486 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1489 if (NextVA.isMemLoc()) {
1490 unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8;
1491 MachineFrameInfo *MFI = MF.getFrameInfo();
1492 int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset());
1494 // Create load node to retrieve arguments from the stack.
1495 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1496 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN, NULL, 0);
1498 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1499 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1502 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, ArgValue, ArgValue2);
1506 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1507 unsigned CallConv, bool isVarArg,
1508 const SmallVectorImpl<ISD::InputArg>
1510 DebugLoc dl, SelectionDAG &DAG,
1511 SmallVectorImpl<SDValue> &InVals) {
1513 MachineFunction &MF = DAG.getMachineFunction();
1514 MachineFrameInfo *MFI = MF.getFrameInfo();
1516 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1518 // Assign locations to all of the incoming arguments.
1519 SmallVector<CCValAssign, 16> ArgLocs;
1520 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1522 CCInfo.AnalyzeFormalArguments(Ins,
1523 CCAssignFnForNode(CallConv, /* Return*/ false,
1526 SmallVector<SDValue, 16> ArgValues;
1528 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1529 CCValAssign &VA = ArgLocs[i];
1531 // Arguments stored in registers.
1532 if (VA.isRegLoc()) {
1533 EVT RegVT = VA.getLocVT();
1536 if (VA.needsCustom()) {
1537 // f64 and vector types are split up into multiple registers or
1538 // combinations of registers and stack slots.
1541 if (VA.getLocVT() == MVT::v2f64) {
1542 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1544 VA = ArgLocs[++i]; // skip ahead to next loc
1545 SDValue ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1547 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1548 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1549 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1550 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1551 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1553 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1556 TargetRegisterClass *RC;
1558 if (RegVT == MVT::f32)
1559 RC = ARM::SPRRegisterClass;
1560 else if (RegVT == MVT::f64)
1561 RC = ARM::DPRRegisterClass;
1562 else if (RegVT == MVT::v2f64)
1563 RC = ARM::QPRRegisterClass;
1564 else if (RegVT == MVT::i32)
1565 RC = (AFI->isThumb1OnlyFunction() ?
1566 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1568 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1570 // Transform the arguments in physical registers into virtual ones.
1571 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1572 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1575 // If this is an 8 or 16-bit value, it is really passed promoted
1576 // to 32 bits. Insert an assert[sz]ext to capture this, then
1577 // truncate to the right size.
1578 switch (VA.getLocInfo()) {
1579 default: llvm_unreachable("Unknown loc info!");
1580 case CCValAssign::Full: break;
1581 case CCValAssign::BCvt:
1582 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1584 case CCValAssign::SExt:
1585 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1586 DAG.getValueType(VA.getValVT()));
1587 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1589 case CCValAssign::ZExt:
1590 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1591 DAG.getValueType(VA.getValVT()));
1592 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1596 InVals.push_back(ArgValue);
1598 } else { // VA.isRegLoc()
1601 assert(VA.isMemLoc());
1602 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1604 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1605 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset());
1607 // Create load nodes to retrieve arguments from the stack.
1608 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1609 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN, NULL, 0));
1615 static const unsigned GPRArgRegs[] = {
1616 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1619 unsigned NumGPRs = CCInfo.getFirstUnallocated
1620 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1622 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1623 unsigned VARegSize = (4 - NumGPRs) * 4;
1624 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1625 unsigned ArgOffset = 0;
1626 if (VARegSaveSize) {
1627 // If this function is vararg, store any remaining integer argument regs
1628 // to their spots on the stack so that they may be loaded by deferencing
1629 // the result of va_next.
1630 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1631 ArgOffset = CCInfo.getNextStackOffset();
1632 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1633 VARegSaveSize - VARegSize);
1634 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1636 SmallVector<SDValue, 4> MemOps;
1637 for (; NumGPRs < 4; ++NumGPRs) {
1638 TargetRegisterClass *RC;
1639 if (AFI->isThumb1OnlyFunction())
1640 RC = ARM::tGPRRegisterClass;
1642 RC = ARM::GPRRegisterClass;
1644 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1645 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1646 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1647 MemOps.push_back(Store);
1648 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1649 DAG.getConstant(4, getPointerTy()));
1651 if (!MemOps.empty())
1652 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1653 &MemOps[0], MemOps.size());
1655 // This will point to the next argument passed via stack.
1656 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
1662 /// isFloatingPointZero - Return true if this is +0.0.
1663 static bool isFloatingPointZero(SDValue Op) {
1664 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1665 return CFP->getValueAPF().isPosZero();
1666 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1667 // Maybe this has already been legalized into the constant pool?
1668 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1669 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1670 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1671 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1672 return CFP->getValueAPF().isPosZero();
1678 static bool isLegalCmpImmediate(unsigned C, bool isThumb1Only) {
1679 return ( isThumb1Only && (C & ~255U) == 0) ||
1680 (!isThumb1Only && ARM_AM::getSOImmVal(C) != -1);
1683 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1684 /// the given operands.
1685 static SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1686 SDValue &ARMCC, SelectionDAG &DAG, bool isThumb1Only,
1688 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1689 unsigned C = RHSC->getZExtValue();
1690 if (!isLegalCmpImmediate(C, isThumb1Only)) {
1691 // Constant does not fit, try adjusting it by one?
1696 if (isLegalCmpImmediate(C-1, isThumb1Only)) {
1697 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1698 RHS = DAG.getConstant(C-1, MVT::i32);
1703 if (C > 0 && isLegalCmpImmediate(C-1, isThumb1Only)) {
1704 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1705 RHS = DAG.getConstant(C-1, MVT::i32);
1710 if (isLegalCmpImmediate(C+1, isThumb1Only)) {
1711 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1712 RHS = DAG.getConstant(C+1, MVT::i32);
1717 if (C < 0xffffffff && isLegalCmpImmediate(C+1, isThumb1Only)) {
1718 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1719 RHS = DAG.getConstant(C+1, MVT::i32);
1726 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1727 ARMISD::NodeType CompareType;
1730 CompareType = ARMISD::CMP;
1735 CompareType = ARMISD::CMPZ;
1738 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1739 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1742 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1743 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1746 if (!isFloatingPointZero(RHS))
1747 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1749 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1750 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1753 static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG,
1754 const ARMSubtarget *ST) {
1755 EVT VT = Op.getValueType();
1756 SDValue LHS = Op.getOperand(0);
1757 SDValue RHS = Op.getOperand(1);
1758 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1759 SDValue TrueVal = Op.getOperand(2);
1760 SDValue FalseVal = Op.getOperand(3);
1761 DebugLoc dl = Op.getDebugLoc();
1763 if (LHS.getValueType() == MVT::i32) {
1765 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1766 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1767 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1770 ARMCC::CondCodes CondCode, CondCode2;
1771 if (FPCCToARMCC(CC, CondCode, CondCode2))
1772 std::swap(TrueVal, FalseVal);
1774 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1775 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1776 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1777 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1779 if (CondCode2 != ARMCC::AL) {
1780 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1781 // FIXME: Needs another CMP because flag can have but one use.
1782 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1783 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1784 Result, TrueVal, ARMCC2, CCR, Cmp2);
1789 static SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG,
1790 const ARMSubtarget *ST) {
1791 SDValue Chain = Op.getOperand(0);
1792 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1793 SDValue LHS = Op.getOperand(2);
1794 SDValue RHS = Op.getOperand(3);
1795 SDValue Dest = Op.getOperand(4);
1796 DebugLoc dl = Op.getDebugLoc();
1798 if (LHS.getValueType() == MVT::i32) {
1800 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1801 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1802 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1803 Chain, Dest, ARMCC, CCR,Cmp);
1806 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1807 ARMCC::CondCodes CondCode, CondCode2;
1808 if (FPCCToARMCC(CC, CondCode, CondCode2))
1809 // Swap the LHS/RHS of the comparison if needed.
1810 std::swap(LHS, RHS);
1812 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1813 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1814 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1815 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1816 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1817 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1818 if (CondCode2 != ARMCC::AL) {
1819 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1820 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1821 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1826 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) {
1827 SDValue Chain = Op.getOperand(0);
1828 SDValue Table = Op.getOperand(1);
1829 SDValue Index = Op.getOperand(2);
1830 DebugLoc dl = Op.getDebugLoc();
1832 EVT PTy = getPointerTy();
1833 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1834 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1835 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1836 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1837 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
1838 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
1839 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
1840 if (Subtarget->isThumb2()) {
1841 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
1842 // which does another jump to the destination. This also makes it easier
1843 // to translate it to TBB / TBH later.
1844 // FIXME: This might not work if the function is extremely large.
1845 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
1846 Addr, Op.getOperand(2), JTI, UId);
1848 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1849 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr, NULL, 0);
1850 Chain = Addr.getValue(1);
1851 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
1852 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1854 Addr = DAG.getLoad(PTy, dl, Chain, Addr, NULL, 0);
1855 Chain = Addr.getValue(1);
1856 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1860 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1861 DebugLoc dl = Op.getDebugLoc();
1863 Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI;
1864 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
1865 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
1868 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1869 EVT VT = Op.getValueType();
1870 DebugLoc dl = Op.getDebugLoc();
1872 Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF;
1874 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
1875 return DAG.getNode(Opc, dl, VT, Op);
1878 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
1879 // Implement fcopysign with a fabs and a conditional fneg.
1880 SDValue Tmp0 = Op.getOperand(0);
1881 SDValue Tmp1 = Op.getOperand(1);
1882 DebugLoc dl = Op.getDebugLoc();
1883 EVT VT = Op.getValueType();
1884 EVT SrcVT = Tmp1.getValueType();
1885 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
1886 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
1887 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
1888 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1889 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
1892 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
1893 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1894 MFI->setFrameAddressIsTaken(true);
1895 EVT VT = Op.getValueType();
1896 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
1897 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1898 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
1899 ? ARM::R7 : ARM::R11;
1900 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
1902 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
1907 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
1909 SDValue Dst, SDValue Src,
1910 SDValue Size, unsigned Align,
1912 const Value *DstSV, uint64_t DstSVOff,
1913 const Value *SrcSV, uint64_t SrcSVOff){
1914 // Do repeated 4-byte loads and stores. To be improved.
1915 // This requires 4-byte alignment.
1916 if ((Align & 3) != 0)
1918 // This requires the copy size to be a constant, preferrably
1919 // within a subtarget-specific limit.
1920 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
1923 uint64_t SizeVal = ConstantSize->getZExtValue();
1924 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
1927 unsigned BytesLeft = SizeVal & 3;
1928 unsigned NumMemOps = SizeVal >> 2;
1929 unsigned EmittedNumMemOps = 0;
1931 unsigned VTSize = 4;
1933 const unsigned MAX_LOADS_IN_LDM = 6;
1934 SDValue TFOps[MAX_LOADS_IN_LDM];
1935 SDValue Loads[MAX_LOADS_IN_LDM];
1936 uint64_t SrcOff = 0, DstOff = 0;
1938 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
1939 // same number of stores. The loads and stores will get combined into
1940 // ldm/stm later on.
1941 while (EmittedNumMemOps < NumMemOps) {
1943 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1944 Loads[i] = DAG.getLoad(VT, dl, Chain,
1945 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
1946 DAG.getConstant(SrcOff, MVT::i32)),
1947 SrcSV, SrcSVOff + SrcOff);
1948 TFOps[i] = Loads[i].getValue(1);
1951 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1954 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1955 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
1956 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
1957 DAG.getConstant(DstOff, MVT::i32)),
1958 DstSV, DstSVOff + DstOff);
1961 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1963 EmittedNumMemOps += i;
1969 // Issue loads / stores for the trailing (1 - 3) bytes.
1970 unsigned BytesLeftSave = BytesLeft;
1973 if (BytesLeft >= 2) {
1981 Loads[i] = DAG.getLoad(VT, dl, Chain,
1982 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
1983 DAG.getConstant(SrcOff, MVT::i32)),
1984 SrcSV, SrcSVOff + SrcOff);
1985 TFOps[i] = Loads[i].getValue(1);
1988 BytesLeft -= VTSize;
1990 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1993 BytesLeft = BytesLeftSave;
1995 if (BytesLeft >= 2) {
2003 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2004 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2005 DAG.getConstant(DstOff, MVT::i32)),
2006 DstSV, DstSVOff + DstOff);
2009 BytesLeft -= VTSize;
2011 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2014 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2015 SDValue Op = N->getOperand(0);
2016 DebugLoc dl = N->getDebugLoc();
2017 if (N->getValueType(0) == MVT::f64) {
2018 // Turn i64->f64 into FMDRR.
2019 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2020 DAG.getConstant(0, MVT::i32));
2021 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2022 DAG.getConstant(1, MVT::i32));
2023 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
2026 // Turn f64->i64 into FMRRD.
2027 SDValue Cvt = DAG.getNode(ARMISD::FMRRD, dl,
2028 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2030 // Merge the pieces into a single i64 value.
2031 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2034 /// getZeroVector - Returns a vector of specified type with all zero elements.
2036 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2037 assert(VT.isVector() && "Expected a vector type");
2039 // Zero vectors are used to represent vector negation and in those cases
2040 // will be implemented with the NEON VNEG instruction. However, VNEG does
2041 // not support i64 elements, so sometimes the zero vectors will need to be
2042 // explicitly constructed. For those cases, and potentially other uses in
2043 // the future, always build zero vectors as <4 x i32> or <2 x i32> bitcasted
2044 // to their dest type. This ensures they get CSE'd.
2046 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2047 if (VT.getSizeInBits() == 64)
2048 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2050 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2052 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2055 /// getOnesVector - Returns a vector of specified type with all bits set.
2057 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2058 assert(VT.isVector() && "Expected a vector type");
2060 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2061 // type. This ensures they get CSE'd.
2063 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2064 if (VT.getSizeInBits() == 64)
2065 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2067 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2069 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2072 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2073 const ARMSubtarget *ST) {
2074 EVT VT = N->getValueType(0);
2075 DebugLoc dl = N->getDebugLoc();
2077 // Lower vector shifts on NEON to use VSHL.
2078 if (VT.isVector()) {
2079 assert(ST->hasNEON() && "unexpected vector shift");
2081 // Left shifts translate directly to the vshiftu intrinsic.
2082 if (N->getOpcode() == ISD::SHL)
2083 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2084 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2085 N->getOperand(0), N->getOperand(1));
2087 assert((N->getOpcode() == ISD::SRA ||
2088 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2090 // NEON uses the same intrinsics for both left and right shifts. For
2091 // right shifts, the shift amounts are negative, so negate the vector of
2093 EVT ShiftVT = N->getOperand(1).getValueType();
2094 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2095 getZeroVector(ShiftVT, DAG, dl),
2097 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2098 Intrinsic::arm_neon_vshifts :
2099 Intrinsic::arm_neon_vshiftu);
2100 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2101 DAG.getConstant(vshiftInt, MVT::i32),
2102 N->getOperand(0), NegatedCount);
2105 // We can get here for a node like i32 = ISD::SHL i32, i64
2109 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2110 "Unknown shift to lower!");
2112 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2113 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2114 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2117 // If we are in thumb mode, we don't have RRX.
2118 if (ST->isThumb1Only()) return SDValue();
2120 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2121 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2122 DAG.getConstant(0, MVT::i32));
2123 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2124 DAG.getConstant(1, MVT::i32));
2126 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2127 // captures the result into a carry flag.
2128 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2129 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2131 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2132 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2134 // Merge the pieces into a single i64 value.
2135 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2138 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2139 SDValue TmpOp0, TmpOp1;
2140 bool Invert = false;
2144 SDValue Op0 = Op.getOperand(0);
2145 SDValue Op1 = Op.getOperand(1);
2146 SDValue CC = Op.getOperand(2);
2147 EVT VT = Op.getValueType();
2148 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2149 DebugLoc dl = Op.getDebugLoc();
2151 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2152 switch (SetCCOpcode) {
2153 default: llvm_unreachable("Illegal FP comparison"); break;
2155 case ISD::SETNE: Invert = true; // Fallthrough
2157 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2159 case ISD::SETLT: Swap = true; // Fallthrough
2161 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2163 case ISD::SETLE: Swap = true; // Fallthrough
2165 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2166 case ISD::SETUGE: Swap = true; // Fallthrough
2167 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2168 case ISD::SETUGT: Swap = true; // Fallthrough
2169 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2170 case ISD::SETUEQ: Invert = true; // Fallthrough
2172 // Expand this to (OLT | OGT).
2176 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2177 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2179 case ISD::SETUO: Invert = true; // Fallthrough
2181 // Expand this to (OLT | OGE).
2185 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2186 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2190 // Integer comparisons.
2191 switch (SetCCOpcode) {
2192 default: llvm_unreachable("Illegal integer comparison"); break;
2193 case ISD::SETNE: Invert = true;
2194 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2195 case ISD::SETLT: Swap = true;
2196 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2197 case ISD::SETLE: Swap = true;
2198 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2199 case ISD::SETULT: Swap = true;
2200 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2201 case ISD::SETULE: Swap = true;
2202 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2205 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2206 if (Opc == ARMISD::VCEQ) {
2209 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2211 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2214 // Ignore bitconvert.
2215 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2216 AndOp = AndOp.getOperand(0);
2218 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2220 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2221 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2228 std::swap(Op0, Op1);
2230 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2233 Result = DAG.getNOT(dl, Result, VT);
2238 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2239 /// VMOV instruction, and if so, return the constant being splatted.
2240 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2241 unsigned SplatBitSize, SelectionDAG &DAG) {
2242 switch (SplatBitSize) {
2244 // Any 1-byte value is OK.
2245 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2246 return DAG.getTargetConstant(SplatBits, MVT::i8);
2249 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2250 if ((SplatBits & ~0xff) == 0 ||
2251 (SplatBits & ~0xff00) == 0)
2252 return DAG.getTargetConstant(SplatBits, MVT::i16);
2256 // NEON's 32-bit VMOV supports splat values where:
2257 // * only one byte is nonzero, or
2258 // * the least significant byte is 0xff and the second byte is nonzero, or
2259 // * the least significant 2 bytes are 0xff and the third is nonzero.
2260 if ((SplatBits & ~0xff) == 0 ||
2261 (SplatBits & ~0xff00) == 0 ||
2262 (SplatBits & ~0xff0000) == 0 ||
2263 (SplatBits & ~0xff000000) == 0)
2264 return DAG.getTargetConstant(SplatBits, MVT::i32);
2266 if ((SplatBits & ~0xffff) == 0 &&
2267 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2268 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2270 if ((SplatBits & ~0xffffff) == 0 &&
2271 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2272 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2274 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2275 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2276 // VMOV.I32. A (very) minor optimization would be to replicate the value
2277 // and fall through here to test for a valid 64-bit splat. But, then the
2278 // caller would also need to check and handle the change in size.
2282 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2283 uint64_t BitMask = 0xff;
2285 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2286 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2288 else if ((SplatBits & BitMask) != 0)
2292 return DAG.getTargetConstant(Val, MVT::i64);
2296 llvm_unreachable("unexpected size for isVMOVSplat");
2303 /// getVMOVImm - If this is a build_vector of constants which can be
2304 /// formed by using a VMOV instruction of the specified element size,
2305 /// return the constant being splatted. The ByteSize field indicates the
2306 /// number of bytes of each element [1248].
2307 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2308 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2309 APInt SplatBits, SplatUndef;
2310 unsigned SplatBitSize;
2312 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2313 HasAnyUndefs, ByteSize * 8))
2316 if (SplatBitSize > ByteSize * 8)
2319 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2323 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2324 bool &ReverseVEXT, unsigned &Imm) {
2325 unsigned NumElts = VT.getVectorNumElements();
2326 ReverseVEXT = false;
2329 // If this is a VEXT shuffle, the immediate value is the index of the first
2330 // element. The other shuffle indices must be the successive elements after
2332 unsigned ExpectedElt = Imm;
2333 for (unsigned i = 1; i < NumElts; ++i) {
2334 // Increment the expected index. If it wraps around, it may still be
2335 // a VEXT but the source vectors must be swapped.
2337 if (ExpectedElt == NumElts * 2) {
2342 if (ExpectedElt != static_cast<unsigned>(M[i]))
2346 // Adjust the index value if the source operands will be swapped.
2353 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2354 /// instruction with the specified blocksize. (The order of the elements
2355 /// within each block of the vector is reversed.)
2356 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2357 unsigned BlockSize) {
2358 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2359 "Only possible block sizes for VREV are: 16, 32, 64");
2361 unsigned NumElts = VT.getVectorNumElements();
2362 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2363 unsigned BlockElts = M[0] + 1;
2365 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2368 for (unsigned i = 0; i < NumElts; ++i) {
2369 if ((unsigned) M[i] !=
2370 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2377 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2378 unsigned &WhichResult) {
2379 unsigned NumElts = VT.getVectorNumElements();
2380 WhichResult = (M[0] == 0 ? 0 : 1);
2381 for (unsigned i = 0; i < NumElts; i += 2) {
2382 if ((unsigned) M[i] != i + WhichResult ||
2383 (unsigned) M[i+1] != i + NumElts + WhichResult)
2389 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2390 unsigned &WhichResult) {
2391 unsigned NumElts = VT.getVectorNumElements();
2392 WhichResult = (M[0] == 0 ? 0 : 1);
2393 for (unsigned i = 0; i != NumElts; ++i) {
2394 if ((unsigned) M[i] != 2 * i + WhichResult)
2398 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2399 if (VT.is64BitVector() && VT.getVectorElementType().getSizeInBits() == 32)
2405 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2406 unsigned &WhichResult) {
2407 unsigned NumElts = VT.getVectorNumElements();
2408 WhichResult = (M[0] == 0 ? 0 : 1);
2409 unsigned Idx = WhichResult * NumElts / 2;
2410 for (unsigned i = 0; i != NumElts; i += 2) {
2411 if ((unsigned) M[i] != Idx ||
2412 (unsigned) M[i+1] != Idx + NumElts)
2417 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2418 if (VT.is64BitVector() && VT.getVectorElementType().getSizeInBits() == 32)
2424 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2425 // Canonicalize all-zeros and all-ones vectors.
2426 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2427 if (ConstVal->isNullValue())
2428 return getZeroVector(VT, DAG, dl);
2429 if (ConstVal->isAllOnesValue())
2430 return getOnesVector(VT, DAG, dl);
2433 if (VT.is64BitVector()) {
2434 switch (Val.getValueType().getSizeInBits()) {
2435 case 8: CanonicalVT = MVT::v8i8; break;
2436 case 16: CanonicalVT = MVT::v4i16; break;
2437 case 32: CanonicalVT = MVT::v2i32; break;
2438 case 64: CanonicalVT = MVT::v1i64; break;
2439 default: llvm_unreachable("unexpected splat element type"); break;
2442 assert(VT.is128BitVector() && "unknown splat vector size");
2443 switch (Val.getValueType().getSizeInBits()) {
2444 case 8: CanonicalVT = MVT::v16i8; break;
2445 case 16: CanonicalVT = MVT::v8i16; break;
2446 case 32: CanonicalVT = MVT::v4i32; break;
2447 case 64: CanonicalVT = MVT::v2i64; break;
2448 default: llvm_unreachable("unexpected splat element type"); break;
2452 // Build a canonical splat for this value.
2453 SmallVector<SDValue, 8> Ops;
2454 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2455 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2457 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2460 // If this is a case we can't handle, return null and let the default
2461 // expansion code take care of it.
2462 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2463 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2464 DebugLoc dl = Op.getDebugLoc();
2465 EVT VT = Op.getValueType();
2467 APInt SplatBits, SplatUndef;
2468 unsigned SplatBitSize;
2470 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2471 if (SplatBitSize <= 64) {
2472 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2473 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2475 return BuildSplat(Val, VT, DAG, dl);
2479 // If there are only 2 elements in a 128-bit vector, insert them into an
2480 // undef vector. This handles the common case for 128-bit vector argument
2481 // passing, where the insertions should be translated to subreg accesses
2482 // with no real instructions.
2483 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2484 SDValue Val = DAG.getUNDEF(VT);
2485 SDValue Op0 = Op.getOperand(0);
2486 SDValue Op1 = Op.getOperand(1);
2487 if (Op0.getOpcode() != ISD::UNDEF)
2488 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2489 DAG.getIntPtrConstant(0));
2490 if (Op1.getOpcode() != ISD::UNDEF)
2491 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2492 DAG.getIntPtrConstant(1));
2499 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2500 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2501 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2502 /// are assumed to be legal.
2504 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2506 if (VT.getVectorNumElements() == 4 &&
2507 (VT.is128BitVector() || VT.is64BitVector())) {
2508 unsigned PFIndexes[4];
2509 for (unsigned i = 0; i != 4; ++i) {
2513 PFIndexes[i] = M[i];
2516 // Compute the index in the perfect shuffle table.
2517 unsigned PFTableIndex =
2518 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2519 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2520 unsigned Cost = (PFEntry >> 30);
2527 unsigned Imm, WhichResult;
2529 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2530 isVREVMask(M, VT, 64) ||
2531 isVREVMask(M, VT, 32) ||
2532 isVREVMask(M, VT, 16) ||
2533 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2534 isVTRNMask(M, VT, WhichResult) ||
2535 isVUZPMask(M, VT, WhichResult) ||
2536 isVZIPMask(M, VT, WhichResult));
2539 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2540 /// the specified operations to build the shuffle.
2541 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2542 SDValue RHS, SelectionDAG &DAG,
2544 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2545 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2546 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2549 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2558 OP_VUZPL, // VUZP, left result
2559 OP_VUZPR, // VUZP, right result
2560 OP_VZIPL, // VZIP, left result
2561 OP_VZIPR, // VZIP, right result
2562 OP_VTRNL, // VTRN, left result
2563 OP_VTRNR // VTRN, right result
2566 if (OpNum == OP_COPY) {
2567 if (LHSID == (1*9+2)*9+3) return LHS;
2568 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2572 SDValue OpLHS, OpRHS;
2573 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2574 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2575 EVT VT = OpLHS.getValueType();
2578 default: llvm_unreachable("Unknown shuffle opcode!");
2580 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
2585 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
2586 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
2590 return DAG.getNode(ARMISD::VEXT, dl, VT,
2592 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
2595 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2596 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
2599 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2600 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
2603 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2604 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
2608 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2609 SDValue V1 = Op.getOperand(0);
2610 SDValue V2 = Op.getOperand(1);
2611 DebugLoc dl = Op.getDebugLoc();
2612 EVT VT = Op.getValueType();
2613 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
2614 SmallVector<int, 8> ShuffleMask;
2616 // Convert shuffles that are directly supported on NEON to target-specific
2617 // DAG nodes, instead of keeping them as shuffles and matching them again
2618 // during code selection. This is more efficient and avoids the possibility
2619 // of inconsistencies between legalization and selection.
2620 // FIXME: floating-point vectors should be canonicalized to integer vectors
2621 // of the same time so that they get CSEd properly.
2622 SVN->getMask(ShuffleMask);
2624 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
2625 int Lane = SVN->getSplatIndex();
2626 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
2627 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
2629 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
2630 DAG.getConstant(Lane, MVT::i32));
2635 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
2638 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
2639 DAG.getConstant(Imm, MVT::i32));
2642 if (isVREVMask(ShuffleMask, VT, 64))
2643 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
2644 if (isVREVMask(ShuffleMask, VT, 32))
2645 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
2646 if (isVREVMask(ShuffleMask, VT, 16))
2647 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
2649 // Check for Neon shuffles that modify both input vectors in place.
2650 // If both results are used, i.e., if there are two shuffles with the same
2651 // source operands and with masks corresponding to both results of one of
2652 // these operations, DAG memoization will ensure that a single node is
2653 // used for both shuffles.
2654 unsigned WhichResult;
2655 if (isVTRNMask(ShuffleMask, VT, WhichResult))
2656 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2657 V1, V2).getValue(WhichResult);
2658 if (isVUZPMask(ShuffleMask, VT, WhichResult))
2659 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2660 V1, V2).getValue(WhichResult);
2661 if (isVZIPMask(ShuffleMask, VT, WhichResult))
2662 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2663 V1, V2).getValue(WhichResult);
2665 // If the shuffle is not directly supported and it has 4 elements, use
2666 // the PerfectShuffle-generated table to synthesize it from other shuffles.
2667 if (VT.getVectorNumElements() == 4 &&
2668 (VT.is128BitVector() || VT.is64BitVector())) {
2669 unsigned PFIndexes[4];
2670 for (unsigned i = 0; i != 4; ++i) {
2671 if (ShuffleMask[i] < 0)
2674 PFIndexes[i] = ShuffleMask[i];
2677 // Compute the index in the perfect shuffle table.
2678 unsigned PFTableIndex =
2679 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2681 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2682 unsigned Cost = (PFEntry >> 30);
2685 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
2691 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
2692 EVT VT = Op.getValueType();
2693 DebugLoc dl = Op.getDebugLoc();
2694 SDValue Vec = Op.getOperand(0);
2695 SDValue Lane = Op.getOperand(1);
2697 // FIXME: This is invalid for 8 and 16-bit elements - the information about
2698 // sign / zero extension is lost!
2699 Op = DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
2700 Op = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Op, DAG.getValueType(VT));
2702 if (VT.bitsLT(MVT::i32))
2703 Op = DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
2704 else if (VT.bitsGT(MVT::i32))
2705 Op = DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op);
2710 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
2711 // The only time a CONCAT_VECTORS operation can have legal types is when
2712 // two 64-bit vectors are concatenated to a 128-bit vector.
2713 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
2714 "unexpected CONCAT_VECTORS");
2715 DebugLoc dl = Op.getDebugLoc();
2716 SDValue Val = DAG.getUNDEF(MVT::v2f64);
2717 SDValue Op0 = Op.getOperand(0);
2718 SDValue Op1 = Op.getOperand(1);
2719 if (Op0.getOpcode() != ISD::UNDEF)
2720 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2721 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
2722 DAG.getIntPtrConstant(0));
2723 if (Op1.getOpcode() != ISD::UNDEF)
2724 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2725 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
2726 DAG.getIntPtrConstant(1));
2727 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
2730 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
2731 switch (Op.getOpcode()) {
2732 default: llvm_unreachable("Don't know how to custom lower this!");
2733 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
2734 case ISD::GlobalAddress:
2735 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
2736 LowerGlobalAddressELF(Op, DAG);
2737 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
2738 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG, Subtarget);
2739 case ISD::BR_CC: return LowerBR_CC(Op, DAG, Subtarget);
2740 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
2741 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
2742 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
2743 case ISD::SINT_TO_FP:
2744 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
2745 case ISD::FP_TO_SINT:
2746 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
2747 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
2748 case ISD::RETURNADDR: break;
2749 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
2750 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
2751 case ISD::INTRINSIC_VOID:
2752 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG);
2753 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2754 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
2757 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
2758 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
2759 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
2760 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
2761 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2762 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
2767 /// ReplaceNodeResults - Replace the results of node with an illegal result
2768 /// type with new values built out of custom code.
2769 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
2770 SmallVectorImpl<SDValue>&Results,
2771 SelectionDAG &DAG) {
2772 switch (N->getOpcode()) {
2774 llvm_unreachable("Don't know how to custom expand this!");
2776 case ISD::BIT_CONVERT:
2777 Results.push_back(ExpandBIT_CONVERT(N, DAG));
2781 SDValue Res = LowerShift(N, DAG, Subtarget);
2783 Results.push_back(Res);
2789 //===----------------------------------------------------------------------===//
2790 // ARM Scheduler Hooks
2791 //===----------------------------------------------------------------------===//
2794 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
2795 MachineBasicBlock *BB) const {
2796 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2797 DebugLoc dl = MI->getDebugLoc();
2798 switch (MI->getOpcode()) {
2800 llvm_unreachable("Unexpected instr type to insert");
2801 case ARM::tMOVCCr_pseudo: {
2802 // To "insert" a SELECT_CC instruction, we actually have to insert the
2803 // diamond control-flow pattern. The incoming instruction knows the
2804 // destination vreg to set, the condition code register to branch on, the
2805 // true/false values to select between, and a branch opcode to use.
2806 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2807 MachineFunction::iterator It = BB;
2813 // cmpTY ccX, r1, r2
2815 // fallthrough --> copy0MBB
2816 MachineBasicBlock *thisMBB = BB;
2817 MachineFunction *F = BB->getParent();
2818 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
2819 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
2820 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
2821 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
2822 F->insert(It, copy0MBB);
2823 F->insert(It, sinkMBB);
2824 // Update machine-CFG edges by first adding all successors of the current
2825 // block to the new block which will contain the Phi node for the select.
2826 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
2827 e = BB->succ_end(); i != e; ++i)
2828 sinkMBB->addSuccessor(*i);
2829 // Next, remove all successors of the current block, and add the true
2830 // and fallthrough blocks as its successors.
2831 while(!BB->succ_empty())
2832 BB->removeSuccessor(BB->succ_begin());
2833 BB->addSuccessor(copy0MBB);
2834 BB->addSuccessor(sinkMBB);
2837 // %FalseValue = ...
2838 // # fallthrough to sinkMBB
2841 // Update machine-CFG edges
2842 BB->addSuccessor(sinkMBB);
2845 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
2848 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
2849 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
2850 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
2852 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
2859 case ARM::t2SUBrSPi_:
2860 case ARM::t2SUBrSPi12_:
2861 case ARM::t2SUBrSPs_: {
2862 MachineFunction *MF = BB->getParent();
2863 unsigned DstReg = MI->getOperand(0).getReg();
2864 unsigned SrcReg = MI->getOperand(1).getReg();
2865 bool DstIsDead = MI->getOperand(0).isDead();
2866 bool SrcIsKill = MI->getOperand(1).isKill();
2868 if (SrcReg != ARM::SP) {
2869 // Copy the source to SP from virtual register.
2870 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
2871 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
2872 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
2873 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
2874 .addReg(SrcReg, getKillRegState(SrcIsKill));
2878 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
2879 switch (MI->getOpcode()) {
2881 llvm_unreachable("Unexpected pseudo instruction!");
2887 OpOpc = ARM::tADDspr;
2890 OpOpc = ARM::tSUBspi;
2892 case ARM::t2SUBrSPi_:
2893 OpOpc = ARM::t2SUBrSPi;
2894 NeedPred = true; NeedCC = true;
2896 case ARM::t2SUBrSPi12_:
2897 OpOpc = ARM::t2SUBrSPi12;
2900 case ARM::t2SUBrSPs_:
2901 OpOpc = ARM::t2SUBrSPs;
2902 NeedPred = true; NeedCC = true; NeedOp3 = true;
2905 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
2906 if (OpOpc == ARM::tAND)
2907 AddDefaultT1CC(MIB);
2908 MIB.addReg(ARM::SP);
2909 MIB.addOperand(MI->getOperand(2));
2911 MIB.addOperand(MI->getOperand(3));
2913 AddDefaultPred(MIB);
2917 // Copy the result from SP to virtual register.
2918 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
2919 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
2920 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
2921 BuildMI(BB, dl, TII->get(CopyOpc))
2922 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
2924 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
2930 //===----------------------------------------------------------------------===//
2931 // ARM Optimization Hooks
2932 //===----------------------------------------------------------------------===//
2935 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
2936 TargetLowering::DAGCombinerInfo &DCI) {
2937 SelectionDAG &DAG = DCI.DAG;
2938 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2939 EVT VT = N->getValueType(0);
2940 unsigned Opc = N->getOpcode();
2941 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
2942 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
2943 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
2944 ISD::CondCode CC = ISD::SETCC_INVALID;
2947 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
2949 SDValue CCOp = Slct.getOperand(0);
2950 if (CCOp.getOpcode() == ISD::SETCC)
2951 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
2954 bool DoXform = false;
2956 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
2959 if (LHS.getOpcode() == ISD::Constant &&
2960 cast<ConstantSDNode>(LHS)->isNullValue()) {
2962 } else if (CC != ISD::SETCC_INVALID &&
2963 RHS.getOpcode() == ISD::Constant &&
2964 cast<ConstantSDNode>(RHS)->isNullValue()) {
2965 std::swap(LHS, RHS);
2966 SDValue Op0 = Slct.getOperand(0);
2967 EVT OpVT = isSlctCC ? Op0.getValueType() :
2968 Op0.getOperand(0).getValueType();
2969 bool isInt = OpVT.isInteger();
2970 CC = ISD::getSetCCInverse(CC, isInt);
2972 if (!TLI.isCondCodeLegal(CC, OpVT))
2973 return SDValue(); // Inverse operator isn't legal.
2980 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
2982 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
2983 Slct.getOperand(0), Slct.getOperand(1), CC);
2984 SDValue CCOp = Slct.getOperand(0);
2986 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
2987 CCOp.getOperand(0), CCOp.getOperand(1), CC);
2988 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
2989 CCOp, OtherOp, Result);
2994 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
2995 static SDValue PerformADDCombine(SDNode *N,
2996 TargetLowering::DAGCombinerInfo &DCI) {
2997 // added by evan in r37685 with no testcase.
2998 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3000 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3001 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3002 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3003 if (Result.getNode()) return Result;
3005 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3006 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3007 if (Result.getNode()) return Result;
3013 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3014 static SDValue PerformSUBCombine(SDNode *N,
3015 TargetLowering::DAGCombinerInfo &DCI) {
3016 // added by evan in r37685 with no testcase.
3017 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3019 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3020 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3021 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3022 if (Result.getNode()) return Result;
3029 /// PerformFMRRDCombine - Target-specific dag combine xforms for ARMISD::FMRRD.
3030 static SDValue PerformFMRRDCombine(SDNode *N,
3031 TargetLowering::DAGCombinerInfo &DCI) {
3032 // fmrrd(fmdrr x, y) -> x,y
3033 SDValue InDouble = N->getOperand(0);
3034 if (InDouble.getOpcode() == ARMISD::FMDRR)
3035 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3039 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3040 /// operand of a vector shift operation, where all the elements of the
3041 /// build_vector must have the same constant integer value.
3042 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3043 // Ignore bit_converts.
3044 while (Op.getOpcode() == ISD::BIT_CONVERT)
3045 Op = Op.getOperand(0);
3046 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3047 APInt SplatBits, SplatUndef;
3048 unsigned SplatBitSize;
3050 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3051 HasAnyUndefs, ElementBits) ||
3052 SplatBitSize > ElementBits)
3054 Cnt = SplatBits.getSExtValue();
3058 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3059 /// operand of a vector shift left operation. That value must be in the range:
3060 /// 0 <= Value < ElementBits for a left shift; or
3061 /// 0 <= Value <= ElementBits for a long left shift.
3062 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3063 assert(VT.isVector() && "vector shift count is not a vector type");
3064 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3065 if (! getVShiftImm(Op, ElementBits, Cnt))
3067 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3070 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3071 /// operand of a vector shift right operation. For a shift opcode, the value
3072 /// is positive, but for an intrinsic the value count must be negative. The
3073 /// absolute value must be in the range:
3074 /// 1 <= |Value| <= ElementBits for a right shift; or
3075 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3076 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3078 assert(VT.isVector() && "vector shift count is not a vector type");
3079 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3080 if (! getVShiftImm(Op, ElementBits, Cnt))
3084 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3087 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3088 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3089 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3092 // Don't do anything for most intrinsics.
3095 // Vector shifts: check for immediate versions and lower them.
3096 // Note: This is done during DAG combining instead of DAG legalizing because
3097 // the build_vectors for 64-bit vector element shift counts are generally
3098 // not legal, and it is hard to see their values after they get legalized to
3099 // loads from a constant pool.
3100 case Intrinsic::arm_neon_vshifts:
3101 case Intrinsic::arm_neon_vshiftu:
3102 case Intrinsic::arm_neon_vshiftls:
3103 case Intrinsic::arm_neon_vshiftlu:
3104 case Intrinsic::arm_neon_vshiftn:
3105 case Intrinsic::arm_neon_vrshifts:
3106 case Intrinsic::arm_neon_vrshiftu:
3107 case Intrinsic::arm_neon_vrshiftn:
3108 case Intrinsic::arm_neon_vqshifts:
3109 case Intrinsic::arm_neon_vqshiftu:
3110 case Intrinsic::arm_neon_vqshiftsu:
3111 case Intrinsic::arm_neon_vqshiftns:
3112 case Intrinsic::arm_neon_vqshiftnu:
3113 case Intrinsic::arm_neon_vqshiftnsu:
3114 case Intrinsic::arm_neon_vqrshiftns:
3115 case Intrinsic::arm_neon_vqrshiftnu:
3116 case Intrinsic::arm_neon_vqrshiftnsu: {
3117 EVT VT = N->getOperand(1).getValueType();
3119 unsigned VShiftOpc = 0;
3122 case Intrinsic::arm_neon_vshifts:
3123 case Intrinsic::arm_neon_vshiftu:
3124 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3125 VShiftOpc = ARMISD::VSHL;
3128 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3129 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3130 ARMISD::VSHRs : ARMISD::VSHRu);
3135 case Intrinsic::arm_neon_vshiftls:
3136 case Intrinsic::arm_neon_vshiftlu:
3137 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3139 llvm_unreachable("invalid shift count for vshll intrinsic");
3141 case Intrinsic::arm_neon_vrshifts:
3142 case Intrinsic::arm_neon_vrshiftu:
3143 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3147 case Intrinsic::arm_neon_vqshifts:
3148 case Intrinsic::arm_neon_vqshiftu:
3149 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3153 case Intrinsic::arm_neon_vqshiftsu:
3154 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3156 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3158 case Intrinsic::arm_neon_vshiftn:
3159 case Intrinsic::arm_neon_vrshiftn:
3160 case Intrinsic::arm_neon_vqshiftns:
3161 case Intrinsic::arm_neon_vqshiftnu:
3162 case Intrinsic::arm_neon_vqshiftnsu:
3163 case Intrinsic::arm_neon_vqrshiftns:
3164 case Intrinsic::arm_neon_vqrshiftnu:
3165 case Intrinsic::arm_neon_vqrshiftnsu:
3166 // Narrowing shifts require an immediate right shift.
3167 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3169 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3172 llvm_unreachable("unhandled vector shift");
3176 case Intrinsic::arm_neon_vshifts:
3177 case Intrinsic::arm_neon_vshiftu:
3178 // Opcode already set above.
3180 case Intrinsic::arm_neon_vshiftls:
3181 case Intrinsic::arm_neon_vshiftlu:
3182 if (Cnt == VT.getVectorElementType().getSizeInBits())
3183 VShiftOpc = ARMISD::VSHLLi;
3185 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3186 ARMISD::VSHLLs : ARMISD::VSHLLu);
3188 case Intrinsic::arm_neon_vshiftn:
3189 VShiftOpc = ARMISD::VSHRN; break;
3190 case Intrinsic::arm_neon_vrshifts:
3191 VShiftOpc = ARMISD::VRSHRs; break;
3192 case Intrinsic::arm_neon_vrshiftu:
3193 VShiftOpc = ARMISD::VRSHRu; break;
3194 case Intrinsic::arm_neon_vrshiftn:
3195 VShiftOpc = ARMISD::VRSHRN; break;
3196 case Intrinsic::arm_neon_vqshifts:
3197 VShiftOpc = ARMISD::VQSHLs; break;
3198 case Intrinsic::arm_neon_vqshiftu:
3199 VShiftOpc = ARMISD::VQSHLu; break;
3200 case Intrinsic::arm_neon_vqshiftsu:
3201 VShiftOpc = ARMISD::VQSHLsu; break;
3202 case Intrinsic::arm_neon_vqshiftns:
3203 VShiftOpc = ARMISD::VQSHRNs; break;
3204 case Intrinsic::arm_neon_vqshiftnu:
3205 VShiftOpc = ARMISD::VQSHRNu; break;
3206 case Intrinsic::arm_neon_vqshiftnsu:
3207 VShiftOpc = ARMISD::VQSHRNsu; break;
3208 case Intrinsic::arm_neon_vqrshiftns:
3209 VShiftOpc = ARMISD::VQRSHRNs; break;
3210 case Intrinsic::arm_neon_vqrshiftnu:
3211 VShiftOpc = ARMISD::VQRSHRNu; break;
3212 case Intrinsic::arm_neon_vqrshiftnsu:
3213 VShiftOpc = ARMISD::VQRSHRNsu; break;
3216 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3217 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3220 case Intrinsic::arm_neon_vshiftins: {
3221 EVT VT = N->getOperand(1).getValueType();
3223 unsigned VShiftOpc = 0;
3225 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3226 VShiftOpc = ARMISD::VSLI;
3227 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3228 VShiftOpc = ARMISD::VSRI;
3230 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3233 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3234 N->getOperand(1), N->getOperand(2),
3235 DAG.getConstant(Cnt, MVT::i32));
3238 case Intrinsic::arm_neon_vqrshifts:
3239 case Intrinsic::arm_neon_vqrshiftu:
3240 // No immediate versions of these to check for.
3247 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3248 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3249 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3250 /// vector element shift counts are generally not legal, and it is hard to see
3251 /// their values after they get legalized to loads from a constant pool.
3252 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3253 const ARMSubtarget *ST) {
3254 EVT VT = N->getValueType(0);
3256 // Nothing to be done for scalar shifts.
3257 if (! VT.isVector())
3260 assert(ST->hasNEON() && "unexpected vector shift");
3263 switch (N->getOpcode()) {
3264 default: llvm_unreachable("unexpected shift opcode");
3267 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3268 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3269 DAG.getConstant(Cnt, MVT::i32));
3274 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3275 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3276 ARMISD::VSHRs : ARMISD::VSHRu);
3277 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3278 DAG.getConstant(Cnt, MVT::i32));
3284 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3285 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3286 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3287 const ARMSubtarget *ST) {
3288 SDValue N0 = N->getOperand(0);
3290 // Check for sign- and zero-extensions of vector extract operations of 8-
3291 // and 16-bit vector elements. NEON supports these directly. They are
3292 // handled during DAG combining because type legalization will promote them
3293 // to 32-bit types and it is messy to recognize the operations after that.
3294 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3295 SDValue Vec = N0.getOperand(0);
3296 SDValue Lane = N0.getOperand(1);
3297 EVT VT = N->getValueType(0);
3298 EVT EltVT = N0.getValueType();
3299 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3301 if (VT == MVT::i32 &&
3302 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3303 TLI.isTypeLegal(Vec.getValueType())) {
3306 switch (N->getOpcode()) {
3307 default: llvm_unreachable("unexpected opcode");
3308 case ISD::SIGN_EXTEND:
3309 Opc = ARMISD::VGETLANEs;
3311 case ISD::ZERO_EXTEND:
3312 case ISD::ANY_EXTEND:
3313 Opc = ARMISD::VGETLANEu;
3316 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3323 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
3324 DAGCombinerInfo &DCI) const {
3325 switch (N->getOpcode()) {
3327 case ISD::ADD: return PerformADDCombine(N, DCI);
3328 case ISD::SUB: return PerformSUBCombine(N, DCI);
3329 case ARMISD::FMRRD: return PerformFMRRDCombine(N, DCI);
3330 case ISD::INTRINSIC_WO_CHAIN:
3331 return PerformIntrinsicCombine(N, DCI.DAG);
3335 return PerformShiftCombine(N, DCI.DAG, Subtarget);
3336 case ISD::SIGN_EXTEND:
3337 case ISD::ZERO_EXTEND:
3338 case ISD::ANY_EXTEND:
3339 return PerformExtendCombine(N, DCI.DAG, Subtarget);
3344 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
3345 if (!Subtarget->hasV6Ops())
3346 // Pre-v6 does not support unaligned mem access.
3348 else if (!Subtarget->hasV6Ops()) {
3349 // v6 may or may not support unaligned mem access.
3350 if (!Subtarget->isTargetDarwin())
3354 switch (VT.getSimpleVT().SimpleTy) {
3361 // FIXME: VLD1 etc with standard alignment is legal.
3365 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
3370 switch (VT.getSimpleVT().SimpleTy) {
3371 default: return false;
3386 if ((V & (Scale - 1)) != 0)
3389 return V == (V & ((1LL << 5) - 1));
3392 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
3393 const ARMSubtarget *Subtarget) {
3400 switch (VT.getSimpleVT().SimpleTy) {
3401 default: return false;
3406 // + imm12 or - imm8
3408 return V == (V & ((1LL << 8) - 1));
3409 return V == (V & ((1LL << 12) - 1));
3412 // Same as ARM mode. FIXME: NEON?
3413 if (!Subtarget->hasVFP2())
3418 return V == (V & ((1LL << 8) - 1));
3422 /// isLegalAddressImmediate - Return true if the integer value can be used
3423 /// as the offset of the target addressing mode for load / store of the
3425 static bool isLegalAddressImmediate(int64_t V, EVT VT,
3426 const ARMSubtarget *Subtarget) {
3433 if (Subtarget->isThumb1Only())
3434 return isLegalT1AddressImmediate(V, VT);
3435 else if (Subtarget->isThumb2())
3436 return isLegalT2AddressImmediate(V, VT, Subtarget);
3441 switch (VT.getSimpleVT().SimpleTy) {
3442 default: return false;
3447 return V == (V & ((1LL << 12) - 1));
3450 return V == (V & ((1LL << 8) - 1));
3453 if (!Subtarget->hasVFP2()) // FIXME: NEON?
3458 return V == (V & ((1LL << 8) - 1));
3462 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
3464 int Scale = AM.Scale;
3468 switch (VT.getSimpleVT().SimpleTy) {
3469 default: return false;
3478 return Scale == 2 || Scale == 4 || Scale == 8;
3481 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3485 // Note, we allow "void" uses (basically, uses that aren't loads or
3486 // stores), because arm allows folding a scale into many arithmetic
3487 // operations. This should be made more precise and revisited later.
3489 // Allow r << imm, but the imm has to be a multiple of two.
3490 if (Scale & 1) return false;
3491 return isPowerOf2_32(Scale);
3495 /// isLegalAddressingMode - Return true if the addressing mode represented
3496 /// by AM is legal for this target, for a load/store of the specified type.
3497 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3498 const Type *Ty) const {
3499 EVT VT = getValueType(Ty, true);
3500 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
3503 // Can never fold addr of global into load/store.
3508 case 0: // no scale reg, must be "r+i" or "r", or "i".
3511 if (Subtarget->isThumb1Only())
3515 // ARM doesn't support any R+R*scale+imm addr modes.
3522 if (Subtarget->isThumb2())
3523 return isLegalT2ScaledAddressingMode(AM, VT);
3525 int Scale = AM.Scale;
3526 switch (VT.getSimpleVT().SimpleTy) {
3527 default: return false;
3531 if (Scale < 0) Scale = -Scale;
3535 return isPowerOf2_32(Scale & ~1);
3539 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3544 // Note, we allow "void" uses (basically, uses that aren't loads or
3545 // stores), because arm allows folding a scale into many arithmetic
3546 // operations. This should be made more precise and revisited later.
3548 // Allow r << imm, but the imm has to be a multiple of two.
3549 if (Scale & 1) return false;
3550 return isPowerOf2_32(Scale);
3557 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
3558 bool isSEXTLoad, SDValue &Base,
3559 SDValue &Offset, bool &isInc,
3560 SelectionDAG &DAG) {
3561 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3564 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
3566 Base = Ptr->getOperand(0);
3567 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3568 int RHSC = (int)RHS->getZExtValue();
3569 if (RHSC < 0 && RHSC > -256) {
3570 assert(Ptr->getOpcode() == ISD::ADD);
3572 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3576 isInc = (Ptr->getOpcode() == ISD::ADD);
3577 Offset = Ptr->getOperand(1);
3579 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
3581 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3582 int RHSC = (int)RHS->getZExtValue();
3583 if (RHSC < 0 && RHSC > -0x1000) {
3584 assert(Ptr->getOpcode() == ISD::ADD);
3586 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3587 Base = Ptr->getOperand(0);
3592 if (Ptr->getOpcode() == ISD::ADD) {
3594 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
3595 if (ShOpcVal != ARM_AM::no_shift) {
3596 Base = Ptr->getOperand(1);
3597 Offset = Ptr->getOperand(0);
3599 Base = Ptr->getOperand(0);
3600 Offset = Ptr->getOperand(1);
3605 isInc = (Ptr->getOpcode() == ISD::ADD);
3606 Base = Ptr->getOperand(0);
3607 Offset = Ptr->getOperand(1);
3611 // FIXME: Use FLDM / FSTM to emulate indexed FP load / store.
3615 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
3616 bool isSEXTLoad, SDValue &Base,
3617 SDValue &Offset, bool &isInc,
3618 SelectionDAG &DAG) {
3619 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3622 Base = Ptr->getOperand(0);
3623 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3624 int RHSC = (int)RHS->getZExtValue();
3625 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
3626 assert(Ptr->getOpcode() == ISD::ADD);
3628 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3630 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
3631 isInc = Ptr->getOpcode() == ISD::ADD;
3632 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
3640 /// getPreIndexedAddressParts - returns true by value, base pointer and
3641 /// offset pointer and addressing mode by reference if the node's address
3642 /// can be legally represented as pre-indexed load / store address.
3644 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
3646 ISD::MemIndexedMode &AM,
3647 SelectionDAG &DAG) const {
3648 if (Subtarget->isThumb1Only())
3653 bool isSEXTLoad = false;
3654 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3655 Ptr = LD->getBasePtr();
3656 VT = LD->getMemoryVT();
3657 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3658 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3659 Ptr = ST->getBasePtr();
3660 VT = ST->getMemoryVT();
3665 bool isLegal = false;
3666 if (Subtarget->isThumb2())
3667 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3668 Offset, isInc, DAG);
3670 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3671 Offset, isInc, DAG);
3675 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
3679 /// getPostIndexedAddressParts - returns true by value, base pointer and
3680 /// offset pointer and addressing mode by reference if this node can be
3681 /// combined with a load / store to form a post-indexed load / store.
3682 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
3685 ISD::MemIndexedMode &AM,
3686 SelectionDAG &DAG) const {
3687 if (Subtarget->isThumb1Only())
3692 bool isSEXTLoad = false;
3693 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3694 VT = LD->getMemoryVT();
3695 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3696 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3697 VT = ST->getMemoryVT();
3702 bool isLegal = false;
3703 if (Subtarget->isThumb2())
3704 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3707 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3712 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
3716 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
3720 const SelectionDAG &DAG,
3721 unsigned Depth) const {
3722 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3723 switch (Op.getOpcode()) {
3725 case ARMISD::CMOV: {
3726 // Bits are known zero/one if known on the LHS and RHS.
3727 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
3728 if (KnownZero == 0 && KnownOne == 0) return;
3730 APInt KnownZeroRHS, KnownOneRHS;
3731 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
3732 KnownZeroRHS, KnownOneRHS, Depth+1);
3733 KnownZero &= KnownZeroRHS;
3734 KnownOne &= KnownOneRHS;
3740 //===----------------------------------------------------------------------===//
3741 // ARM Inline Assembly Support
3742 //===----------------------------------------------------------------------===//
3744 /// getConstraintType - Given a constraint letter, return the type of
3745 /// constraint it is for this target.
3746 ARMTargetLowering::ConstraintType
3747 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
3748 if (Constraint.size() == 1) {
3749 switch (Constraint[0]) {
3751 case 'l': return C_RegisterClass;
3752 case 'w': return C_RegisterClass;
3755 return TargetLowering::getConstraintType(Constraint);
3758 std::pair<unsigned, const TargetRegisterClass*>
3759 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3761 if (Constraint.size() == 1) {
3762 // GCC RS6000 Constraint Letters
3763 switch (Constraint[0]) {
3765 if (Subtarget->isThumb1Only())
3766 return std::make_pair(0U, ARM::tGPRRegisterClass);
3768 return std::make_pair(0U, ARM::GPRRegisterClass);
3770 return std::make_pair(0U, ARM::GPRRegisterClass);
3773 return std::make_pair(0U, ARM::SPRRegisterClass);
3775 return std::make_pair(0U, ARM::DPRRegisterClass);
3779 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3782 std::vector<unsigned> ARMTargetLowering::
3783 getRegClassForInlineAsmConstraint(const std::string &Constraint,
3785 if (Constraint.size() != 1)
3786 return std::vector<unsigned>();
3788 switch (Constraint[0]) { // GCC ARM Constraint Letters
3791 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3792 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3795 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3796 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3797 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
3798 ARM::R12, ARM::LR, 0);
3801 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
3802 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
3803 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
3804 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
3805 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
3806 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
3807 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
3808 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
3810 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
3811 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
3812 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
3813 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
3817 return std::vector<unsigned>();
3820 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3821 /// vector. If it is invalid, don't add anything to Ops.
3822 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3825 std::vector<SDValue>&Ops,
3826 SelectionDAG &DAG) const {
3827 SDValue Result(0, 0);
3829 switch (Constraint) {
3831 case 'I': case 'J': case 'K': case 'L':
3832 case 'M': case 'N': case 'O':
3833 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3837 int64_t CVal64 = C->getSExtValue();
3838 int CVal = (int) CVal64;
3839 // None of these constraints allow values larger than 32 bits. Check
3840 // that the value fits in an int.
3844 switch (Constraint) {
3846 if (Subtarget->isThumb1Only()) {
3847 // This must be a constant between 0 and 255, for ADD
3849 if (CVal >= 0 && CVal <= 255)
3851 } else if (Subtarget->isThumb2()) {
3852 // A constant that can be used as an immediate value in a
3853 // data-processing instruction.
3854 if (ARM_AM::getT2SOImmVal(CVal) != -1)
3857 // A constant that can be used as an immediate value in a
3858 // data-processing instruction.
3859 if (ARM_AM::getSOImmVal(CVal) != -1)
3865 if (Subtarget->isThumb()) { // FIXME thumb2
3866 // This must be a constant between -255 and -1, for negated ADD
3867 // immediates. This can be used in GCC with an "n" modifier that
3868 // prints the negated value, for use with SUB instructions. It is
3869 // not useful otherwise but is implemented for compatibility.
3870 if (CVal >= -255 && CVal <= -1)
3873 // This must be a constant between -4095 and 4095. It is not clear
3874 // what this constraint is intended for. Implemented for
3875 // compatibility with GCC.
3876 if (CVal >= -4095 && CVal <= 4095)
3882 if (Subtarget->isThumb1Only()) {
3883 // A 32-bit value where only one byte has a nonzero value. Exclude
3884 // zero to match GCC. This constraint is used by GCC internally for
3885 // constants that can be loaded with a move/shift combination.
3886 // It is not useful otherwise but is implemented for compatibility.
3887 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
3889 } else if (Subtarget->isThumb2()) {
3890 // A constant whose bitwise inverse can be used as an immediate
3891 // value in a data-processing instruction. This can be used in GCC
3892 // with a "B" modifier that prints the inverted value, for use with
3893 // BIC and MVN instructions. It is not useful otherwise but is
3894 // implemented for compatibility.
3895 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
3898 // A constant whose bitwise inverse can be used as an immediate
3899 // value in a data-processing instruction. This can be used in GCC
3900 // with a "B" modifier that prints the inverted value, for use with
3901 // BIC and MVN instructions. It is not useful otherwise but is
3902 // implemented for compatibility.
3903 if (ARM_AM::getSOImmVal(~CVal) != -1)
3909 if (Subtarget->isThumb1Only()) {
3910 // This must be a constant between -7 and 7,
3911 // for 3-operand ADD/SUB immediate instructions.
3912 if (CVal >= -7 && CVal < 7)
3914 } else if (Subtarget->isThumb2()) {
3915 // A constant whose negation can be used as an immediate value in a
3916 // data-processing instruction. This can be used in GCC with an "n"
3917 // modifier that prints the negated value, for use with SUB
3918 // instructions. It is not useful otherwise but is implemented for
3920 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
3923 // A constant whose negation can be used as an immediate value in a
3924 // data-processing instruction. This can be used in GCC with an "n"
3925 // modifier that prints the negated value, for use with SUB
3926 // instructions. It is not useful otherwise but is implemented for
3928 if (ARM_AM::getSOImmVal(-CVal) != -1)
3934 if (Subtarget->isThumb()) { // FIXME thumb2
3935 // This must be a multiple of 4 between 0 and 1020, for
3936 // ADD sp + immediate.
3937 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
3940 // A power of two or a constant between 0 and 32. This is used in
3941 // GCC for the shift amount on shifted register operands, but it is
3942 // useful in general for any shift amounts.
3943 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
3949 if (Subtarget->isThumb()) { // FIXME thumb2
3950 // This must be a constant between 0 and 31, for shift amounts.
3951 if (CVal >= 0 && CVal <= 31)
3957 if (Subtarget->isThumb()) { // FIXME thumb2
3958 // This must be a multiple of 4 between -508 and 508, for
3959 // ADD/SUB sp = sp + immediate.
3960 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
3965 Result = DAG.getTargetConstant(CVal, Op.getValueType());
3969 if (Result.getNode()) {
3970 Ops.push_back(Result);
3973 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,